《经济学人》劝华盛顿：让中国抄美国AI

US Treasury Secretary Bessent said the US could sanction China over "this kind of theft".

By theft, he means Chinese labs copying American companies' models.

The ones making the demand are two American companies, Anthropic and OpenAI.

They want the government to block this path and punish the companies involved.

This kind of copying is called "distillation" in the industry, and the method is not complicated.

First, have a large and expensive model answer a huge number of questions.

Then use those answers to train a much smaller model.

The model trained this way is much cheaper and much faster.

Both companies prohibit this practice in their terms of service.

But the terms only apply to users who have signed contracts; they cannot control outside companies.

So Washington began to act, and Senator Hagerty proposed a bill.

Under this bill, the US could sanction foreign companies engaged in "model extraction".

It could also put them on the entity list, making it hard for American companies to do business with them.

In September this year, the US accused six Chinese companies of copying American AI products.

China responded that these accusations are "baseless".

OpenAI said some users linked to Moonshot AI tried to extract its reasoning capabilities.

This activity once surged, with about 16,000 requests appearing.

More than 4,000 users made requests, and by late July the activity was blocked.

Anthropic, meanwhile, took its case before the Senate Banking Committee.

It said Alibaba's behavior was "the largest known distillation attack".

During the same period, Anthropic agreed to pay about 1.5 billion dollars.

This money was to settle a class-action lawsuit brought by authors.

However, some people think Washington should not actually ban distillation.

The first reason they give is time.

Experts estimate that Chinese labs are only three to six months behind the most frontier models.

Put another way, letting China keep copying is actually beneficial to the US.

Accustomed to copying, Chinese labs always wait for the US to produce something first and then catch up.

So a ban could only buy the US a few months of time.

The second reason has to do with copyright.

These companies usually say that training models on public data counts as "fair use".

But things written by machines are not themselves protected by copyright.

The same logic, turned around, also works against them.

The Soviet Union once built chips from stolen American blueprints, and in the end still fell behind.

So getting something copied into your hands does not mean you can catch up.

In the view of these people, distillation would instead lock China into second place.

This bill is still in the Senate.

Whether Washington will really act depends on what Congress does next.